Source note: Local copy of a Summer Into AI 2026 submission originally published on Substack. Screenshots, video, and audio found in the post body are mirrored locally when publicly accessible; profile and avatar images are intentionally not copied. Read the canonical post.
Boomtown was ambitious. Too ambitious. You can read about the week 2 theme here.
What I aimed for:
After talking through some ideas with GPT 5.5, I felt it had a pretty clear understanding of what I wanted. It rendered an image that really got me excited about possibilities. The art style really jumped out at me, though, I want it to be real.
WOW! If that were an actual playable game that had well balanced mechanics, maybe even some 2 player co-op, that’d be a fun Friday night!
That’s a lot of textures and geometry. I…uh…(checks watch) have to deliver this by Sunday night.
What I actually got done:
I decided that a procedural approach would give me something I could control and dial into to taste faster than waiting for AI to give me a real 3d implementation of something I was passionate about.
This did not end up being nearly as well polished as I like to make things. It’s barely playable…if at all, but this was time-boxed to 1 week, so I’m still happy with the results.
I was far too involved in the process and should have spent more time making tools for the agents. Instead, I had it make me tools so I could drive.
The goal was inspired by Starcraft. Marquee select your workers, right click a resource building (water tower, lumber mill, quarry, fireworks factory), right click a delivery building (fireworks factory, fireworks launch pad), the workers will mine resource and deliver them to the target building. Work through 4 day cycles, build as many fireworks as you can and ready the launch pad before July 4th at 8pm.
The island environment is procedurally generated with an impressive amount of controls.
You can place the buildings anywhere on the island you want. Placing them on trees clears them out of the way. The buildings rotation on 45deg increments too. Fun for town planning!
Once all buildings are placed, clicking the paths button will auto generate path textures in between the buildings which become the strict walking routes for the workers.
Tech Stack:
Vanilla HTML, CSS, Typescript, and Vite
Three.js 3D engine
Dev Pipeline:
Codex for coding, fixing my gripes, etc.
Linear app for work ticketing
Chat GPT for 3D asset generation.
Things I learned:
Playwright MCP is very token hungry. A picture is worth a thousand words after all, so I suppose this makes sense.
GPT 5.5 is very powerful. It can one shot .glb model files surprisingly well. I took screenshots of each house, or factory, or building in the artistic rendering it made (seen top of article) and asked it to explain it’s style in great detail. After it wrote a few paragraphs, I simply asked it to create the .glb file, and it did! This worked surprisingly well! I’m tempted to think it was because it thought through all the details of the image, wrote them all out and then made the model.
Retrying the last bullet point with enough detail on an animation engine for a bipedal walk cycle that is traversed by a normalized input, ChatGPT 5.5 one shot a worker walk cycle with a js driver. It got the normals backwards on the workers hardhat… but….wow.
Public Repo: here
Playable Demo: here
These two buttons are important for starting out:
Original source
Canonical Substack URL: https://kylesebestyen.substack.com/p/boomtown.